Josef Fruehwald (2017)
Virtually all analysis of sound change in progress has focused on how averages change over time (e.g. mean normalized F1), with relatively less quantitative focus on how variances may change (e.g. normalized F1 standard deviation). Linear mixed effects models do provide estimates of inter-speaker variance (Johnson 2009; Drager & Hay 2012) but these have rarely been utilized for inference, and are not typically able to vary according to other parameters like speakers' gender and date of birth. However, both inter and intra-speaker variance are of potential interest for the study of sociolinguistics. Shifts in intra-speaker variance could be indicative of the importance of outliers (Labov, Baranowski & Dinkin 2010) or shifts in speakers' stylistic range (Tamminga, Ahern & Ecay 2016). If inter-speaker variance decreases over the course of a change, that could be evidence of the community converging on a new norm from innovative speakers or listeners (e.g. Barker, Archangeli & Mielke 2011).
In this paper, I estimate how the averages as well as the intra and inter-speaker variances change for two vowel shifts in Philadelphia, drawn from the Philadelphia Neighborhood Corpus, using a non-linear model implemented in Stan. The two changes (/ay0/ and /aw/ raising) were chosen to contrast the gender effect on means. /ay0/ raising is very slightly male led, while /aw/ raising is very strongly female led (Labov 2001; Labov et al 2013).
The procedure for developing inferences from variances is not as well developed as that for averages, but some things are clear from these results. First, despite the fact that these two changes are very different with respect to the gender effect on means, they're very similar in the gender effect on variances. Second, it would seem that men and women have very similar intra-speaker variances across both changes. One possible interpretation of this is that they have similar stylistic ranges around their individual means that remain constant across all stages of these changes. Finally, the smaller range of inter-speaker variance for men means there is a narrower range of variation between men than between women. This could be understood in terms the greater importance for women to adopt flamboyant personae (Eckert 2011), while the decrease in men's inter-speaker variance for /ay/ from the 1960s onwards could coincide with that variable's association with toughness (Conn 2005; Wagner 2008).
These results largely affirm the soundness of relying on averages to analyze sound change, but some interesting nuances can be found in modelling the variance parameters as well.
Presented at NWAV46